Research On Image Dehazing Algorithm Based On Deep Learning

With the continuous development of video surveillance,drone aerial photography,and remote sensing monitoring,image processing technology has attracted more and more attention from scholars.In the haze weather,suspended particles in the air will cause light to scatter,causing energy attenuation,causing problems such as deterioration of the image obtained by imaging equipment,and affecting people's normal study and life.Therefore,the research of image defogging technology is of great significance.Because of the existing types of defogging algorithms,the defogging algorithm based on deep learning has stronger generalization ability and better defogging effect.Therefore,this paper mainly studies this type of algorithm.The specific research work is as follows:Aiming at the problem that the current defogging algorithm does not adequately process the different frequency domains of the haze image and the defogging effect is not good,a deep learning defogging algorithm based on multi-scale segmentation is proposed.In this algorithm,based on the image segmentation method,the haze image is decomposed into four sub-images of low frequency,medium and low frequency,medium and high frequency and high frequency,and a transmittance generation network model composed of four sub-network channels of different complexity is constructed.Then the image fusion technology is added to the model,and after processing,the transmittance map corresponding to the haze image is obtained.Then use the obtained transmittance map to find the most suitable pixel and estimate the atmospheric light value.Finally,the restored clear image is obtained according to the atmospheric scattering model.Through experiments on the synthetic haze image and the real haze image,as well as the comparison of the defogging results with other defogging algorithms,it can be proved that this algorithm can effectively defog the haze image under complex background,and the image after defogging with high contrast and no distortion,it is more suitable for human observation and subsequent processing.At present,most defogging algorithms are based on the atmospheric scattering model,but the defogging effect is often poor due to the estimation error of the variable values in the model.Therefore,a generation anti-haze algorithm based on gradient penalty Wasserstein is proposed.The algorithm achieves end-to-end defogging,and the output clear image is directly obtained from the input haze image.The generating network part of this algorithm adopts a U-Net-like network,which is composed of an encoder,a decoder and a jump connection structure,and a Dense Block module is added to the network to better extract image features.The discriminant network part is composed of multi-layer convolution and fully connected structure.In the loss function part,in addition to adding content-aware loss,a gradient penalty term is added to optimize the loss of the entire network.Through experiments on the synthetic haze image and the real haze image,as well as the comparison of the defogging results with other defogging algorithms,the results prove that this defogging algorithm can effectively remove the haze information in the haze image and enhance the edge detail texture as well as color contrast,the defogging ability has obvious advantages.